Image Classification Algorithm Based On Salient Segmentation

Image classification is of great importance for image understanding and pattern recognition. Due to variety of image content, the visual semantic information is usually objective decided and task-oriented, which is a great challenge to blind image classification. However, blind image without prior information could be classified depending on the guidance of human vision perception. Since the success in visual attention model, Itti’s theory has extensive attention. Therefore, the blind image classification task consist of three parts:1) salient regions extraction for natural images; 2) salient object segmentation; 3) segmented object classification. The main contributions of this paper are listed as follows.(1) A salient object segmentation algorithm is proposed. Motivated by salient blocks are rare visually in the similar images set. Firstly, the method achieves the similar image set of the target image using Gist features. And it combines Scale-invariant feature, SURF and Lab color space for feature extraction. Then, it conducts the similarity measure using Euclidean distance and gets the attention value. Combined with the results of regions marked obtained through the efficient image segmentation method based on graph theory, the method achieves accurate detection of attention regions.(2) An image classification based on multiple instance learning is presented in this paper, combining with visual vocabulary. By extracting SURF descriptor from the segmented image regions, the SURF features will be vectorized by K-means. Then a similarity matrix between each bag and the visual words in the dictionary will be computed, and each image instance will be mapped to produce a visual word dictionary. Finally the 1-norm SVM method is applied to classify images. The experimental results show that the efficiency and effectiveness of the presented approach.In this thesis, we apply and extend the saliency extraction method in the application of the image classification. The proposed methods could get satisfactory results without prior information about the natural image, which can enrich the algorithms and applications of the effective and efficient image annotation and retrieval.